A miracle of miniaturisation
The honeybee, like other insects, is a miracle of miniaturisation and precision-engineering. However, unlike many other insects, the honeybee is also particularly intelligent; honeybees live in large colonies and therefore have to interact socially and, once they become foragers, navigate many kilometres to and from foraging patches - learning routes, landmarks and other associations. It has even been known for honeybees to learn concepts such as 'sameness' and 'difference'. All of this is achieved by a brain of only one million neurons occupying approximately a cubic millimetre. With these one million neurons, the honeybee can achieve even more impressive behaviours than the comparatively simple workhorse of neuroscience and genetics - the fruit fly. Fruit flies have only around 100,000 neurons and do not live in a colony, so their navigational and learning abilities are comparatively much weaker.
Building better robot brains
For several years my lab has been pursuing EPSRC-funded research to try and reverse engineer the honeybee brain, and translate some of its abilities into computational models that we can use to control autonomous flying robots. Flying robots, especially small ones, pose an interesting engineering problem. While much work is focused on driverless cars, this gives a comparatively simple challenge; the high payload of a car allows for multiple redundant and sophisticated sensors and a lot of computational power to process all the data. Driverless cars also work in a comparatively constrained environment, in that currently they are limited to the earth's surface. In contrast, existing flying robots are extremely constrained by current battery performance and thus have tiny payloads for sensors and computation; they also must navigate in potentially complex 3D spaces. Because of this, it seems particularly appropriate to turn to the very efficient bee brain for inspiration.
Evolution has already learned
One of our earliest successes has been to develop a model of the honeybee optic circuits that reproduce honeybee behaviours of speed regulation and obstacle avoidance, all based on estimating optic flow across the robot's camera. Our approach is very robust to variation in the environment. Often, observers assume that we have trained a neural network in a variety of environments, using deep learning; however our model is not trained, rather it is a description of how honeybee neurons are wired together. By reverse engineering the honeybee brain we take advantage of the millions of years of evolution that have tuned the brain circuits of honeybees and other flying insects; further training is not required.
Flying robots - present and future
In our latest EPSRC-funded research we are now making use of the latest developments in mobile computing technology based on 3D computer graphics hardware accelerators. The weight and energy requirements of these have now reached a point where we can mount them on-board small quadcopters, running our neural simulations to give 'brains on board' for true autonomy. However, both battery life and miniaturisation remain substantial challenges for the future. The Harvard's Robobees project, for example, has developed bee-sized flying robots but without on-board power or computation. The challenge of building robots that truly capture the abilities of honeybees, including longevity, size, and behavioural flexibility, will be one for the future.